Human-Aligned Evaluation of a Pixel-wise DNN Color Constancy Model

Researchers have developed a Deep Neural Network (DNN) that matches human color constancy performance in virtual reality. Instead of using physical ground truth, the model was evaluated using the same achromatic object selection task as human subjects. The ResNet-based U-Net showed a strong correspondence with human behavior, achieving high constancy under baseline conditions and mirroring performance declines when key visual cues were removed, demonstrating near-human perceptual alignment.